Human Disease Blood Atlas Olink Data Quality Control

1 Data overview

This section provides an overview of the Olink dataset, including the total number of samples and proteins measured.

In the delivered Olink dataset, there are currently 2018 samples, and 5416 proteins measured in each sample.

Figure 1: Number of samples per cohort.

Code

data_hdba |> 
  filter(!SampleType %in% control_sample_types) |> 
  distinct(DAid) |> 
  left_join(manifest, by = "DAid") |> 
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  count(Cohort) |> 
  ggplot(aes(Cohort, n, fill = Cohort)) +
  geom_col() +
    scale_fill_manual(values = brewer.pal(n = 11, name = "Set3")) +  
  geom_text(aes(label = n), vjust = -0.5) +
  theme_hpa(angled = T) +
  ylab("Number of samples")

Code

#ggsave(savepath("da_phase2_cohorts.png"), h = 6, w = 6)

Figure 2: Sample distribution across plates.

Code

data_hdba |> 
  mutate(Plate = gsub("P", "", Plate),
         Plate = as.numeric(Plate),
         Plate = as.factor(Plate)) |>
  distinct(DAid, SampleID, Plate, SampleType) |> 
  filter(!is.na(Plate)) |> 
  left_join(manifest, by = "DAid") |> 
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_"),
         Cohort = ifelse(SampleType %in% control_sample_types, "CONTROLS", Cohort))  |> 
  group_by(Plate, Cohort) |> 
  count() |> 
  ggplot(aes(Plate, n, fill = Cohort)) +
  geom_col() +
  scale_fill_manual(values = brewer.pal(n = 11, name = "Set3")) +  
  theme_hpa(angled = T) +
  ylab("Number of samples")

Code

#ggsave(savepath("da_phase2_plates.png"), h = 6, w = 8)

2 Step 1: Exploration of control samples and control assays

This section identifies and analyzes control samples to ensure data quality. Specifically, we explore correlations between technical replicates for selected patients (patient 2 and patient 3) across multiple plates.

2.1 Plate-to-plate correlation for patient 2 & 3

We calculate Spearman correlations for measurements of Patient 2 and Patient 3 across plates. These analyses help assess technical reproducibility.

Figure 3: Correlation heatmap for Patient 2 across plates.

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12362") |> 
  select(Plate, OlinkID, PCNormalizedNPX) |> 
  pivot_wider(names_from = OlinkID, values_from = PCNormalizedNPX) |> 
  column_to_rownames("Plate") |> 
  t() |> 
  cor(method = "spearman")|> 
  pheatmap()|> 
  as.ggplot()

Code

#ggsave(savepath("p2_heatmap.png"), h = 8, w = 8)

Figure 4: Correlation heatmap for Patient 3 across plates.

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12363") |> 
  select(Plate, OlinkID, NPX) |> 
  pivot_wider(names_from = OlinkID, values_from = NPX) |> 
  column_to_rownames("Plate") |> 
  t() |> 
  cor(method = "spearman")|> 
  pheatmap() |> 
  as.ggplot()

Code

#ggsave(savepath("p3_heatmap.png"), h = 8, w = 8)

Figure 5: Scatterplots showing pairwise correlations between plates for Patient 2.

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12362") |> 
  select(Plate, OlinkID, NPX) |> 
  pivot_wider(names_from = Plate, values_from = NPX) |> 
  ggpairs(c(2:13),
            lower = list(continuous = "points", size = 0.05)) +
  ggtitle("Correlation for Patient 2 measurements across plates") 

Code

#ggsave(savepath("p2_plate_scatterplots.png"), h = 10, w = 10)

Figure 6: Scatterplots showing pairwise correlations between plates for Patient 3.

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12363") |> 
  select(Plate, OlinkID, NPX) |> 
  pivot_wider(names_from = Plate, values_from = NPX) |> 
  ggpairs(c(2:13),
            lower = list(continuous = "points", size = 0.5)) +
  ggtitle("Correlation for Patient 3 measurements across plates")

Code

#ggsave(savepath("p3_plate_scatterplots.png"), h = 10, w = 10)

2.1.1 Correlation excluding proteins below LOD

This section refines the correlation analysis by excluding proteins that fall below the LOD across all replicates. This step ensures a cleaner analysis focusing only on reliably measured proteins.

Code

exclude_proteins <- 
  data_ht_lod |> 
  filter(DAid %in% control_samples) |> 
  mutate(under_LOD = ifelse(NPX < LOD, "Yes", "No")) |> 
  group_by(Assay) |> 
  count(under_LOD) |> 
  filter(under_LOD == "Yes") |> 
  ungroup() |> 
  filter(n > length(unique(data_ht$PlateID)))

Figure 7: Heatmaps of correlations for Patient 2 measurements (filtered for proteins above LOD).

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12362",
         !Assay %in% exclude_proteins$Assay) |> 
  select(Plate, OlinkID, PCNormalizedNPX) |> 
  pivot_wider(names_from = OlinkID, values_from = PCNormalizedNPX) |> 
  column_to_rownames("Plate") |> 
  t() |> 
  cor(method = "spearman")|> 
  pheatmap() |> 
  as.ggplot()

Code

#ggsave(savepath("p2_subset_heatmap.png"), h = 8, w = 8)

Figure 8: Heatmaps of correlations for Patient 3 measurements (filtered for proteins above LOD).

Code

data_ht |> 
  left_join(id_mapping, by = "SampleID") |> 
  filter(DAid == "DA12363",
         !Assay %in% exclude_proteins$Assay) |> 
  select(Plate, OlinkID, NPX) |> 
  pivot_wider(names_from = OlinkID, values_from = NPX) |> 
  column_to_rownames("Plate") |> 
  t() |> 
  cor(method = "spearman")|> 
  pheatmap() |> 
  as.ggplot()

Code

#ggsave(savepath("p3_subset_heatmap.png"), h = 8, w = 8)

2.2 Remove control assays & samples

In this step, we filter out control assays and samples that are not part of the main dataset. This ensures clean data for downstream analyses.

Code

data_step_1 <- 
  data_ht_lod |> 
  filter(!SampleType %in% control_sample_types,
         !DAid %in% control_samples, # Remove internal KTH controls
         !AssayType %in% control_assay_types) |>
  select(DAid, Assay, OlinkID, UniProt, Panel, Block, Plate, Count, ExtNPX, NPX, PCNormalizedNPX, AssayQC, SampleQC, LOD) |> 
  mutate(above_LOD = ifelse(NPX > LOD, "Yes", "No"))

3 Step 2: QC warnings

In this step, we assess quality control (QC) warnings at the sample and protein levels. Samples or protein measurements flagged with warnings will be excluded based on predefined thresholds.

3.1 Percentage of QC warnings per sample

We calculate the number of QC warnings per sample, with a focus on those that do not pass QC.

Figure 9: Barplot showing the number of QC warnings for each sample, colored by cohort.

Code

data_step_1 |> 
  group_by(DAid) |> 
  count(SampleQC) |> 
  filter(!SampleQC %in% c("NA", "PASS")) |> 
  arrange(-n) |> 
  left_join(manifest, by = "DAid") |> 
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(fct_reorder(DAid, -n), n, fill = Cohort)) +
  geom_col() +
  scale_fill_manual(values = brewer.pal(n = 11, name = "Set3")) +  
  geom_hline(yintercept = 5440/2, linetype = "dashed", color = "grey") +
  theme_hpa(angled = T) +
  xlab("Sample") +
  theme(axis.text.x = element_text(size = 5),
        axis.ticks.x = element_blank()) +
  ggtitle("Samples with QC warnings") 

Code

#ggsave(savepath("samples_qc.png"), h = 4, w = 10)

data_step_1 |> 
  group_by(DAid) |> 
  count(SampleQC) |> 
  filter(!SampleQC %in% c("NA", "PASS")) |> 
  arrange(-n) |> 
  left_join(manifest, by = "DAid") |> 
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(fct_reorder(DAid, -n), n, fill = SampleQC)) +
  geom_col() +
  geom_hline(yintercept = 5440/2, linetype = "dashed", color = "grey") +
  theme_hpa(angled = T) +
  xlab("Sample") +
  theme(axis.text.x = element_text(size = 5),
        axis.ticks.x = element_blank()) +
  ggtitle("Samples with QC warnings") 

Figure 10: Beeswarm plot showing the number of proteins passing QC for each sample.

Code

data_step_1 |> 
  group_by(DAid) |> 
  filter(SampleQC == "PASS") |>
  count(SampleQC) |> 
  left_join(manifest, by = "DAid") |> 
  filter(Cohort != "CTRL") |> 
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(Cohort, n, color = Cohort)) +
  geom_quasirandom() +
  geom_hline(yintercept = n_proteins / 2, color = "grey40", linetype = "dashed") +
 # geom_text_repel(aes(label = DAid), show.legend = F) +
  scale_color_manual(values = pal_phase2) +
  theme_hpa(angled = T) +
  ggtitle("Number of proteins that pass QC per sample")

Code

#ggsave(savepath("samples_pass.png"), h = 8, w = 8)

3.2 Samples with > 50% QC warnings

This analysis identifies samples with more than 50% of proteins flagged with QC warnings, which will be excluded from further analysis.

Code

samples_high_warnings <- 
  data_step_1 |> 
  group_by(DAid) |> 
  count(SampleQC) |> 
  filter(SampleQC == "FAIL") |> 
  mutate(perc_warning = n / n_proteins) |> 
  filter(perc_warning > 0.5) 

In the current dataset, the number of samples with > 50% warnings is: 2.

These samples will be excluded from further analysis.

3.3 Remove individual sample QC warnings

In this step, we filter out the samples identified in the previous step that have more than 50% QC warnings. Additionally, we remove protein measurements that failed QC. Since the data is in long format, this step will not introduce any NAs. However, when the data is reshaped into a wide format, missing values may be introduced.

Code

data_step_2 <- 
  data_step_1 |> 
  filter(SampleQC == "PASS",
         !DAid %in% samples_high_warnings$DAid) |>
  select(-SampleQC)

4 Step 3: Assay warnings

In this step, we evaluate quality control (QC) warnings at the assay level. Assays with a high percentage of warnings will be flagged and excluded from the dataset based on a predefined threshold to ensure data reliability.

Figure 11: Bar plot showing the number of assays with warnings across different assays.

Code

data_step_2 |>
  group_by(Assay) |> 
  count(AssayQC) |> 
  filter(!AssayQC %in% c("NA", "PASS")) |> 
  arrange(-n) |> 
  ggplot(aes(fct_reorder(Assay,-n), n)) +
  geom_col() +
  geom_text(aes(label = n), vjust = -0.5) +
  theme_hpa(angle = T) +
  xlab("Assay")

Code

#ggsave(savepath("assay_qc.png"), h = 5, w = 10)

assays_high_warnings <- 
  data_step_2 |> 
  filter(!DAid %in% control_samples) |> 
  group_by(Assay) |> 
  mutate(n_total = n_distinct(DAid)) |> 
  filter(AssayQC == "WARN") |> 
  mutate(n = n_distinct(DAid)) |> 
  ungroup() |> 
  distinct(Assay, n, n_total) |> 
  mutate(perc_warning = n / n_total) |> arrange(-n) |> 
  filter(perc_warning > 0.5) 

In the current dataset, the number of assays with > 50% warnings is: 1.

These assays will be excluded from further analysis.

Code

data_step_3 <- 
  data_step_2 |> 
  filter(!Assay %in% assays_high_warnings$Assay) |>
  select(-AssayQC)

5 Step 4: Outlier exploration

In this step, we use PCA and UAMP to explore potential outliers in the dataset. By visualizing the data in lower dimensions, we can identify samples that may behave differently from the rest and may need further investigation or exclusion.

Code

# Prepare data
input_data <- 
  data_step_3 |>
  select(DAid, OlinkID, NPX) |>
  pivot_wider(names_from = "OlinkID", values_from = "NPX")

# Exclude protens below LOD in 50% of control measurements
exclude_proteins <- 
  data_ht_lod |> 
  filter(DAid %in% control_samples) |> 
  mutate(under_LOD = ifelse(NPX < LOD, "Yes", "No")) |>
  group_by(Assay) |> 
  count(under_LOD) |> 
  filter(under_LOD == "Yes") |> 
  ungroup() |> 
  filter(n >= sum(!is.na(unique(data_ht_lod$Plate))))

input_data_filt <- 
  data_step_3 |>
  filter(!Assay %in% exclude_proteins$Assay) |> 
  select(DAid, OlinkID, NPX) |>
  pivot_wider(names_from = "OlinkID", values_from = "NPX")

5.1 PCA

Figure 15: PCA visualization based on all proteins, where samples are colored by cohort.

Code

pca_all_proteins <-
  do_pca(data = input_data,
          plots = F)

pca_all_proteins$pca_res |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(PC1, PC2, color = Class)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = pal_class_2) +
  theme_hpa() 

Code

#ggsave(savepath("pca.png"), h = 8, w = 8)

# ids <- 
#   pca_all_proteins$pca_res  |> 
#   filter(PC2 < -70)

# manifest |> 
#   filter(DAid %in% ids$Sample) |> 
#   select(`PI - sample owner`,  Cohort)

We identify the PCA outliers and remove them from the dataset.

Code

pca_outliers <-
  pca_all_proteins$pca_res |>
  filter(PC2 < -70) |>
  distinct(Sample)

data_step_4 <-
  data_step_3 |>
  filter(!DAid %in% pca_outliers$Sample)

Figure 16: PCA visualization based on all proteins and number of proteins detected per cohort, highlighting ourliers.

Code

plot_pca <- 
  pca_all_proteins$pca_res |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Outlier = ifelse(DAid %in% pca_outliers$Sample, "Yes", "No")) |>
  ggplot(aes(PC1, PC2, color = Outlier)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = c("Yes" = "red", "No" = "grey")) +
  theme_hpa() 

missing_per_sample <- 
  data_ht_lod |> 
  filter(SampleType == "SAMPLE") |> 
  mutate(above_LOD = ifelse(NPX > LOD, "Yes", "No")) |> 
  group_by(DAid) |> 
  count(above_LOD) |> 
  filter(above_LOD == "Yes") |> 
  arrange(n)

plot_n <- 
  missing_per_sample |> 
  left_join(manifest, by = "DAid") |> 
  filter(!Cohort %in% c("BDG2", "CTRL")) |> 
  mutate(Outlier = ifelse(DAid %in% pca_outliers$Sample, "Yes", "No")) |>
  ggplot(aes(Cohort, n)) +
  geom_quasirandom(aes(color = Outlier)) +
  geom_boxplot(outlier.shape = NA, alpha = 0.5) +
  scale_color_manual(values = c("grey", "red")) +
  theme_hpa() +
  ggtitle("Number of detected proteins per sample")

plot_pca + plot_n 

Code

#ggsave(savepath("pca_outliers.png"), h = 5, w = 12)

Figure 17: PCA visualization based on proteins with high detectability, where samples are colored by cohort.

Code

pca_high_detectability <-
  do_pca(data = input_data_filt,
         plots = F)

pca_high_detectability$pca_res |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Outlier = ifelse(DAid %in% pca_outliers$Sample, "Yes", "No")) |>
  ggplot(aes(PC1, PC2, color = Outlier)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = c("grey", "red")) +
  theme_hpa() +
  ggtitle("Proteins with high detectability")

Code

#ggsave(savepath("pca_high_detectability.png"), h = 8, w = 8)

5.2 UMAP

Figure 18: UMAP visualization based on all proteins, where samples are colored by cohort.

Code

umap_all_proteins <-
  do_umap(data = input_data,
          plots = F)

umap_all_proteins |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(UMAP1, UMAP2, color = Class)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = pal_class_2) +
  theme_hpa() +
  ggtitle("All proteins")

Code

umap_outliers <-
  umap_all_proteins |> 
  filter(UMAP2 > 5) |> 
  rename(DAid = Sample) |> 
  left_join(manifest, by = "DAid") 

Figure 19: UMAP visualization based on all proteins and number of proteins detected per cohort, highlighting ourliers.

Code

plot_umap <- 
  umap_all_proteins |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Outlier = ifelse(DAid %in% umap_outliers$DAid, "Yes", "No")) |>
  ggplot(aes(UMAP1, UMAP2, color = Outlier)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = c("Yes" = "red", "No" = "grey")) +
  theme_hpa() 

plot_n <- 
  missing_per_sample |> 
  left_join(manifest, by = "DAid") |> 
  filter(!Cohort %in% c("BDG2", "CTRL")) |> 
  mutate(Outlier = ifelse(DAid %in% umap_outliers$DAid, "Yes", "No")) |>
  ggplot(aes(Cohort, n)) +
  geom_quasirandom(aes(color = Outlier)) +
  geom_boxplot(outlier.shape = NA, alpha = 0.5) +
  scale_color_manual(values = c("grey", "red")) +
  theme_hpa() +
  ggtitle("Number of detected proteins per sample")

plot_umap + plot_n 

Code

#ggsave(savepath("umap_outliers.png"), h = 5, w = 12)

Figure 20: UMAP visualization based on proteins with high detectability, where samples are colored by cohort.

Code

umap_high_detectability <-
  do_umap(data = input_data_filt,
          plots = F)

umap_high_detectability |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Outlier = ifelse(DAid %in% umap_outliers$DAid, "Yes", "No")) |>  
  ggplot(aes(UMAP1, UMAP2, color = Outlier)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = c("Yes" = "red", "No" = "grey")) +
  theme_hpa() +
  ggtitle("Proteins with high detectability") 

Code

#ggsave(savepath("umap_high_detectability.png"), h = 8, w = 8)

Code

# Final check PSC1
# 
# umap_high_detectability |>
#   rename(DAid = Sample) |>
#   left_join(manifest, by = "DAid") |>
#   mutate(Outlier = ifelse(DAid %in% umap_outliers$DAid, "Yes", "No")) |>  
#   ggplot(aes(UMAP1, UMAP2, color = Outlier)) +
#   geom_point(alpha = 0.7, size = 0.8) +
#   scale_color_manual(values = c("Yes" = "red", "No" = "grey")) +
#   theme_hpa() +
#   ggtitle("Proteins with high detectability") +
# 
# umap_high_detectability |>
#   rename(DAid = Sample) |>
#   left_join(manifest, by = "DAid") |>
#   ggplot(aes(UMAP1, UMAP2, color = Cohort)) +
#   geom_point(alpha = 0.7, size = 0.8) +
#   scale_fill_manual(values = brewer.pal(n = 11, name = "Set3")) +  
#   theme_hpa() +
#   ggtitle("Proteins with high detectability")
# 
# 
# data_umap_highdetectability <- 
#   data_step_5 |>
#   filter(!Assay %in% exclude_proteins$Assay)
# 
# data_umap_highdetectability
# 
# missing_per_sample <- 
#   data_umap_highdetectability |> 
#   mutate(above_LOD = ifelse(NPX > LOD, "Yes", "No")) |> 
#   group_by(DAid) |> 
#   count(above_LOD) |> 
#   filter(above_LOD == "Yes") |> 
#   arrange(n)
# 
# #plot_n <- 
# missing_per_sample |> 
#   left_join(manifest) |> 
#   filter(!Cohort %in% c("BDG2", "CTRL")) |> 
#   mutate(Outlier_PCA = ifelse(DAid %in% pca_outliers$Sample, "Yes", "No")) |>
#   ggplot(aes(Cohort, n)) +
#   geom_quasirandom(aes(color = Outlier_PCA)) +
#   geom_boxplot(outlier.shape = NA, alpha = 0.5) +
#   scale_color_manual(values = c("grey", "red")) +
#   theme_hpa() +
#   ggtitle("Number of detected proteins per sample")
# 
# new_umap_outliers <- 
#   umap_high_detectability |> 
#   filter(UMAP2 < 2.5, UMAP1 < -8)
# 
# umap_high_detectability |>
#   rename(DAid = Sample) |>
#   left_join(manifest, by = "DAid") |>
#   mutate(Outlier = ifelse(DAid %in% new_umap_outliers$Sample, "Yes", "No")) |>  
#   ggplot(aes(UMAP1, UMAP2, color = Outlier)) +
#   geom_point(alpha = 0.7, size = 0.8) +
#   scale_color_manual(values = c("Yes" = "red", "No" = "grey")) +
#   theme_hpa() +
#   ggtitle("Proteins with high detectability") +
# 
# missing_per_sample |> 
#   left_join(manifest) |> 
#   filter(!Cohort %in% c("BDG2", "CTRL")) |> 
#   mutate(Outlier_new_UMAP = ifelse(DAid %in% new_umap_outliers$Sample, "Yes", "No")) |>
#   ggplot(aes(Cohort, n)) +
#   geom_quasirandom(aes(color = Outlier_new_UMAP)) +
#   geom_boxplot(outlier.shape = NA, alpha = 0.5) +
#   scale_color_manual(values = c("grey", "red")) +
#   theme_hpa() +
#   ggtitle("Number of detected proteins per sample")

6 Final UMAP

Finally, we look at a UMAP of the final dataset after outlier exclusion.

Figure 20: UMAP visualization based on the final dataset.

Code

umap_final <- 
  do_umap(data = data_step_4,
          wide = F,
          plots = F)

umap_final |>
  rename(DAid = Sample) |>
  left_join(manifest, by = "DAid") |>
  mutate(Cohort = paste(Cohort, `PI - sample owner`, sep = "_")) |>
  ggplot(aes(UMAP1, UMAP2, color = Cohort)) +
  geom_point(alpha = 0.7, size = 0.8) +
  scale_color_manual(values = brewer.pal(n = 11, name = "Set3")) +  
  theme_hpa() +
  ggtitle("Final UMAP")

Code

# umap_final |> filter(UMAP1 < -4.3) |> 
#   rename(DAid = Sample) |> 
#   left_join(manifest)
# 
# umap_final |>
#   rename(DAid = Sample) |>
#   left_join(manifest, by = "DAid") |>
#   mutate(Disease = ifelse(Cohort == "PSC1", Disease, NA)) |> 
#   ggplot(aes(UMAP1, UMAP2, color = Disease)) +
#   geom_point(alpha = 0.7, size = 0.8) +
#   #scale_color_manual(values = brewer.pal(n = 11, name = "Set3")) +  
#   theme_hpa() +
#   ggtitle("Final UMAP")

7 Save final data

After completing the quality control and data curation steps, we save the cleaned and processed dataset for sharing with collaborators and for further downstream analysis. This final dataset is ready for use in subsequent phases of the project.

Code

curated_data <- 
  data_step_4 |> 
  select(-Count,    -ExtNPX, -PCNormalizedNPX) # Remove unnecessary columns

write_tsv(curated_data, file = "../data/final_data/data_phase2_batch4_curated_20250425.tsv")